Pattern identification of movement related states in biosignals

The advancement in biosignal processing and modelling has led to exploring the human brain and developing assistive Human Machine Interface (HMI) as well as Brain Machine Interface (BMI). HMI and BMI require specialised techniques for signal processing and pattern recognition to reliably translate information from complex non-stationary dynamics of biosignals into controlling commands. The information translation process consists of signal pre-processing, feature identification and classification. Even though there is continuous progress in biosignal processing research, the critical requirement for HMI and BMI has raised significant challenges for current state-of-art translation methods, such as high accuracy, reliability, and robustness in noise, provided that only small amount of data is available in practice. Therefore, analysing biosignals with novel feature enhancement, feature selection and classification methods are important for decoding of movement intention towards development of reliable assistive HMI as well as BMI. It is particularly valuable for neural signal analysis to understand the neural circuit mechanisms. This research project aims to design decoding algorithm with improved classification performance in robustness and accuracy to recognise movement related states from tongue movement ear pressure (TMEP) signals and deep brain local field potentials (LFPs) by integrating features extracted through multiple domains, and applying pattern classification methods. To achieve the above aim, this project addresses a number of research issues by utilising conventional and efficient signal information extraction, selection and pattern classification techniques. The first part of this research project successfully developed a robust decoding technique for identifying tongue movement commands from TMEP signals in adverse environment for designing an assistive HMI. This decoding strategy utilised wavelet method for optimal feature enhancement and achieved high accuracy in real time with pattern classification methods of Bayesian and support vector machine (SVM). In the second part, the movement commands are decoded from deep brain local field potentials (LFPs) from basal ganglia (Subthalamic Nucleus (STN) or Globus Pallidus interna (GPi)). An efficient translation algorithm is developed to decode deep brain LFPs for identification of movement activities. Neural synchronisation measures including event related desynchronisation and synchronisation, and functional coupling are utilised to extract discriminatory information as features. We further developed a new feature selection strategy named as weighted sequential feature selection (WSFS) to select an optimal feature subset, which is proved robust for high dimensional, small size dataset. Together with WSFS and pattern classification methods (Bayesian or SVM) high decoding performance for identifying movements was achieved. This research work not only assists decoding movement activities for the application of BMI, but also may help to advance understanding of the neural circuit mechanisms related to motor control as well as development of more efficient therapeutic techniques for neuromotor diseases, such as Parkinson disease.